Differences with torch.nn.functional.dropout3d

torch.nn.functional.dropout3d

torch.nn.functional.dropout3d(input, p=0.5, training=True, inplace=False) -> Tensor

For more information, see torch.nn.functional.dropout3d.

mindspore.ops.dropout3d

mindspore.ops.dropout3d(input, p=0.5, training=True) -> Tensor

For more information, see mindspore.ops.dropout3d.

Differences

API function of MindSpore is consistent with that of PyTorch, with differences in the supported data types for parameters.

PyTorch: During training, dropout3d randomly zeroes some channels of the input tensor with probability p from a Bernoulli distribution, each channel will be zeroed out independently on every forward call which based on Bernoulli distribution probability p. Zeroing some channels of the input tensor is proved that it can effectively reduce over fitting and prevent neuronal coadaptation.

MindSpore: Mindspore only supports a tensor with a rank of 5 as input.

Categories	Subcategories	PyTorch	MindSpore	Difference
Parameters	Parameter 1	input	input	Mindspore only supports a tensor with a rank of 5 as input
	Parameter 2	p	p	-
	Parameter 3	training	training	-
	Parameter 4	inplace	-	-

Code Example 1

# PyTorch
import torch

input = torch.ones(2, 3, 2, 4)
output = torch.nn.functional.dropout3d(input)
print(output.shape)
# torch.Size([2, 3, 2, 4])

# MindSpore
import mindspore as ms
from mindspore import ops
from mindspore import Tensor
import numpy as np

input = Tensor(np.ones([2, 3, 2, 4]), ms.float32)
input = input.expand_dims(0)
output = ops.dropout3d(input)
output = output.squeeze(0)
print(output.shape)
# (2, 3, 2, 4)

Code Example 2

# PyTorch
import torch

input = torch.ones(1, 1, 2, 3, 2, 4)
output = torch.nn.functional.dropout3d(input)
print(output.shape)
# torch.Size([1, 1, 2, 3, 2, 4])

# MindSpore
import mindspore as ms
from mindspore import ops
from mindspore import Tensor
import numpy as np

input = Tensor(np.ones([1, 1, 2, 3, 2, 4]), ms.float32)
input = input.squeeze(0)
output = ops.dropout3d(input)
output = output.expand_dims(0)
print(output.shape)
# (1, 1, 2, 3, 2, 4)